Slop Filter v0.6.5

Added

• Reddit support: feed posts, opened posts, and comments on www.reddit.com. A flagged comment folds its own text and leaves its replies visible.
• A tell for the insight reframe: "the hard part is...", "the real work is...", "what most people miss...". It runs on every post and carries a heavy starting weight.
• Hover a post's percentage to see why it got that score: each check's pull, strongest first, and any reply-only questions that were not asked.
• Live stats panel for demos: posts scanned, slop caught, last and average request time, questions answered, tokens, total cost, and cost per 1,000 posts. Off by default, on the options page.
• A privacy policy, and a demo clip at the top of the README.

Changed

• The score is a small chip instead of a full row under every post, so scoring a post no longer makes it taller. The AI and Human label buttons appear when the chip is hovered. A collapsed post still gets the full row.
• Each site places the chip where it has room. On X it sits in the header row. On Reddit comments it sits inline, to the right of Reddit's own badges under the username.
• Posts of 5 or more words are scored. The cutoff was 8, which let one and two phrase slop through.
• The options page saves each setting the moment it changes. The Save button is gone.

Slop Filter v0.6.4

Changed

• New extension icon, by HideMaru from Flaticon, credited in the README, the options page, and assets/icons/ATTRIBUTION.txt.

Slop Filter v0.6.3

Added

• Extension icons, made from the options page mascot. The Chrome Web Store requires them.

Slop Filter v0.6.2

Added

• MIT license, CONTRIBUTING.md, and SKILL.md, which teaches a coding agent to install, update, tune, and change the extension.
• This changelog, release tags, and a GitHub Action that publishes a release with an extension zip when a version tag is pushed.

Changed

• The qualifier tell is now narrower: calling something "real" when no fake or surface version of it was mentioned. A contrast that was set up first, and fixed terms like "real estate", do not count.
• The README install prompt is one sentence that points an agent at SKILL.md.

Removed

• The filler adverb word list. It flagged words people use all the time.

Slop Filter v0.6.1

Added

• A reply-only tell: does the reply turn the post into a general lesson that would fit under almost any post on the topic.
• A tell for emphasis words that add no meaning.
• A filler adverb word list (removed again in 0.6.2).

Slop Filter v0.6.0

Added

• LinkedIn support: feed posts and comments. A comment is judged against the post it sits under.
• Per-site adapters in src/sites/. The core is site-neutral, so a new site is one small file.
• Tells common on LinkedIn: "thrilled to announce" framing, emoji bullet lists, and a pile of hashtags at the end.
• Labels record their site, and scripts/fit.mjs can fit the weights for one site.
• A redesigned options page: bundled Fredoka font, flat colors, light and dark.

Changed

• The state sent to Jev calls the text post.text instead of tweet.text.

Slop Filter v0.5.2

Changed

• Renamed from X AI Filter to Slop Filter.

Slop Filter v0.5.1

The first version in this repository. Versions 0.1.0 to 0.5.0 were development builds from before the repository existed.

Added

• Scores every post on X with TypeSafe Jev: a set of narrow questions per post, combined into one probability with logistic weights.
• Four ways to treat a flagged post: Collapse, Animated, Dim, and Badge.
• Animated mode: a scan line bounces over the post while Jev decides, then the post turns green and stays, or turns red and closes in on itself.
• AI and Human label buttons on every scored post, with export, and scripts/fit.mjs to fit the weights from the labels.
• Retries with exponential backoff, then an Upstream API error line in the bar in place of the score.
• An options page with the API key, threshold, flag mode, weights, a Reload extension button, and a warning when Chrome is running older code than the files on disk.